← @bgSponge

Bend

Summary

Bend uses Lean-style proofs and an affine dependent type theory to block AI coding mistakes and enable massive parallelism via CUDA.

  1. Bend prevents AI errors via proofs.

    Users declare requirements in LAWS.bend, making it mathematically impossible for an AI to merge code that breaks those theorems. [23]

    Merging a bug is mathematically impossible: it is a theorem.

  2. The type checker enables agentic workflows.

    Unlike Lean or Rocq which take minutes, Bend's proof checker finishes in one second, allowing AI agents to verify every change instantly. [10]

    Bend takes a second at most, so an AI agent can check after every change.

  3. Parallelism requires no manual synchronization.

    The runtime eliminates threads, locks, and kernels by automatically spreading calls across all available CPU cores or GPU cores. [12]

    No threads, no locks, no kernels to write.

  4. Native execution matches C performance.

    Bend compiles to native code that performs nearly as fast as C on a single core and scales 100x on a GPU. [8]

  5. Bend favors backend environments.

    The language is currently optimized for Linux and macOS back-end workloads. [30]

Highlights

  • a fast language that blocks AI mistakes via proof

  • Split the work in two, and Bend spreads the calls over every core it can find, then joins them back.

  • Without LAWS.bend, the bug went live. With LAWS.bend, the AI had to retry until it built a wall and proved the law holds.

Read the original ↗Save to your Sponge

Summarized with Sponge. Quotes are the author’s own words.

More from @bg

All readings →